Virtuous Distortion: Abstraction and Idealization in Model-Based Science

“There are the obvious idealizations of physics – infinite potentials, zero time correlations, perfectly rigid rods, and frictionless planes. But it would be a mistake to think entirely in terms of idealization, of properties which we conceive as limiting cases, to which we can approach closer and closer in reality. For some properties are not even approached in reality.”

Nancy Cartwright

I MODELS

1. Distortion

The use of models in the construction of scientific theories is as widespread as it is philosophically interesting (and, one might say, vexing).1 In neither philosophical nor scientific practice do we find a univocal concept of model.2 But there is one established usage to which we want to direct our particular attention in this paper, in which a model is constituted by the theorist’s idealizations and abstractions. Idealizations are expressed by statements known to be false. Abstractions are achieved by suppressing what is known to be true. Idealizations over-represent empirical phenomena. Abstractions under-represent them. We might think of idealizations and abstractions as one another’s duals. Either way, they are purposeful distortions of phenomena on the ground.3

Biologists investigate the action of natural selection in populations by representing them as infinitely large. With stochastic effects safely ignored, changes in gene frequencies can be tracked by concentrating solely on the action of selection.4 Physicists investigate mechanical movement in frictionless planes, by means of which they are able to investigate the simplest kind of movement: rectilinear motion.5 Neoclassical economists work out the laws of supply and demand and downwards

1 See, for example, the online issue of Synthese, 172, 2 (2009).2 For the notion of model-based science see, for example, Peter Godfrey-Smith “The strategy of model based science”, Biology and Philosophy, 21 (2006), 725-740. For model-based reasoning see, for example, Lorenzo Magnani, Nancy Nersession and Paul Thagard, editors, Model-Based Reasoning in Scientific Discovery, Dordrecht/New York: Kluwer/Plenum, 1999, and Lorenzo Magnani, editor, Model Based Reasoning in Science and Engineering, London: College Publications, 2006.3 An early discussion of idealization as distortion is Ernan McMullin’s “Galilean idealization”, Studies in the History and Philosophy of Science, 16, (1985), 247-273. A recent discussion is Michael Weisberg’s “Three kinds of idealization”, Journal of Philosophy, (2007), 639, 659. Weisberg characterizes idealization as “the intentional introduction of distortion into scientific theories” (p. 639)”, and remarks that “idealization should be seen as an activity that involves distorting theories or models, not simply a property of the theory-world relationship” (p. 640). The link between false models and true theories is also a central preoccupation of William Wimsatt’s Re-Engineering Philosophy for Limited Beings: Piecewise Approximations to Reality. Cambridge: Harvard University Press, 2007.4 A good foundational discussion of mathematical models of natural selection can be found in Sean H. Rice, Evolutionary Theory: Mathematical and Conceptual Foundations Suderland: MA, Sinauer Press, 2004.5 See Roberto Torretti, The Philosophy of Physics New York: Cambridge University Press, 1999.

1

inelasticity in models where utilities are infinitely divisible and all motivational factors are suppressed, save for the optimization of subjective expected utility. Likewise, rational decision theorists elucidate the processes of decision-making in models in which, agent-motivation is the maximization of self-interest and nothing more and decisional agents are logically omniscient.6 In the first kind of case, theorists impose something that is false. In the second, they omit things that are true. In the third and fourth, they do both.

There is a widespread and deeply entrenched view among scientists and philosophers that the distortions effected by idealization and abstraction are both necessary and virtuous. Many people who hold this view see the value of models as merely instrumental. Others take the stronger and more philosophically challenging position that the good done by these aspects of scientific modelling has a genuinely cognitive character. The distinction between instrumental and cognitive (or alethic) value is less an algorithmically exact one than an attractive expository convenience.7 Roughly speaking, a theory has instrumental value when it works predictively or helps control events in new or better ways, without necessarily generating new knowledge. A theory has cognitive value when, whatever the character of its other virtues, it enlarges our knowledge of its analytical targets. (For example, particle physics has a good predictive record, but it also brings us a knowledge of how particles behave.) So, then, a short way of making the present claim is this: For wide ranges of cases, saying what’s false and suppressing what is true is indispensable to the cognitive good of science.

Notwithstanding the sheer volume of traffic in the modelling literature, focussed discussions of what makes these distortions facilitators of cognitive success attracts comparatively slight analytical attention by philosophers of science and philosophically-minded scientists. This is perhaps less true of the distortions effected by abstraction than those constituted by idealization. Still, in relation to the scale of use of the models methodology, these discussions aren’t remotely as widespread and, even when they do occur, are not particularly “thick”. The principal purpose of this paper is to thicken our understanding of the cognitive virtuosity of falsehood-promotion and truth-suppression.

The distinction we’ve drawn between idealization and abstraction is amply present in the contemporary literature, although not always with the clarity it both deserves and admits of. In McMullin’s 1985 paper, “Galilean Idealization”,8 the distortions effected by idealization are simplifications facilitate “at least a partial understanding” of the objects modelled. At the risk of confusion, McMullin specifies a further form of distortion which he also identifies with idealization, but which

6 Such idealizations are usually supposed to have normative significance; that is, they are descriptively wrong but normatively authoritative. We ourselves have reservations about this idea, which we lack the space to develop here. Interested readers might wish to consult Dov Gabbay and John Woods, “Normative models of rational agency: The disutility of some approaches”, Logic Journal of the IGPL, 11 (2003), 597-613.7 On the instrumentalist side, recent writers include Ronald Laymon, “Idealiation, explanation and confirmation”, Proceedings of the Biennial Meeting of the Philosophy of Science Association, Vol. 1, (1980), pp. 336-350 and Leslie Nowack, “The idealizational approach to science: A survey”, in J. Brezinski and Leslie Nowack, editors, Idealization III: Idealization and Truth, Amsterdam: Rodopi 1992, pp. 9-63. On the cognitive side, see for example Stephan Hartmann “Idealization in quantum field theory”, in Niall Shanks, editor, (1998), Robert Batterman, “Idealization and modelling”, Synthese, 169 (2009), 427-446, Michael Weisburg, “Three kinds of idealization”, Journal of Philosophy, 104 (2008), 639-659, and Michael Strevens, Depth, Cambridge: Harvard University Press, 2008. 8 Studies in the History and Philosophy of Science 16 (1985), 247-273.

2

corresponds to abstraction in our sense; this involves screening out factors from a complex situation with a view to achieving a better command of the simplified residue. (p. 264). In 1989, Cartwright anticipates our own distinction.9 Idealizations change properties of objects, perfecting them in certain ways, not all of which even approximate to reality. Abstraction on the other hand, is not a matter of “changing any particular features or properties, but rather of subtracting, not only the concrete circumstances but even the material in which the cause in embedded and all that follows from that.” (p. 187)10

George Gale11 draws this same contrast. Abstraction is a process “in which some features are chosen to be represented, and some rejected for representation .” (p. 167). In abstraction, “there is a loss of features when one goes from the subject to its model”. (p. 168) Idealization is a “smoothing out” of the properties represented in the model, in which “the subject feature is ‘perfected’, with no net loss of features”. (p. 168)

Anjan Chakravartty12 like-mindedly proposes that abstraction is a process in which only some of the potentially many relevant factors or parameters present in reality are built-in to a model concerned with a particular class of phenomena”. (p. 327) Abstraction is a matter of ignoring “other parameters that are potentially relevant to the phenomena at issue”, whereas the “hallmark of idealization is that model elements are constructed in such a way as to differ from the systems we take to be their subject matter, not merely by excluding certain parameters, but by enjoying assumptions that could never obtain”. (p. 328)

More recently, Martin R. Jones13 also makes the present distinction, and takes idealization to require the “assertion of a falsehood” and abstraction the “omission of a truth” ( p. 175); and Peter Godfrey-Smith (2009)14 sees idealization as “[t]reating things and having features they clearly do not have”, and abstraction as “[l]eaving things out, while still giving a literally true description”. (2009, p.47)

In their present characterizations, we are introduced to idealization and abstraction in their “pure” forms. Idealization in its pure form is the addition of something false without the suppression of anything true; and abstraction in its pure form is the suppression of something true without the addition of anything false. It bears repeating, however, that in actual practice idealization and abstraction often not only co-occur, but do so in intimately linked ways. As Gale points out, “while this difference

9 Nature’s Capacities and Their Measurement, New York: Cambridge University Press, 1989.10 In “Abstraction via generic modelling in concept formation in science”, in Martin R. Jones and Nancy Cartwright, editors, 2005, Nancy Nersessian resists Cartwright’s distinction, and finds it more “salient” to speak of various abstractive processes, one of which is idealization in Cartwright’s sense, and another is abstraction in Cartwright’s sense. A third is what Nersessian calls “generic modeling.” (p. 137)11 “Idealization and cosmology: A case study”, in Neill Shanks, editor, Idealization IX: Idealization in Contemporary Physics, Amsterdam/New York, Rodopi, 1998, pp. 165-182.12 “The semantic or model-theoretic view of theories and scientific realism”, Synthese 127, (2001), 325-345.13 “Idealization and abstraction: A framework”, in Idealization XII: Idealization and Abstraction in the Sciences, Martin R. Jones and Nancy Cartwright, editors, Amsterdam/New York: Rodopi, 2005, pp. 173-217.14 “Abstractions, idealizations, and evolutionary biology”, in A. Barberousse, M. Morange, and T. Pradeu, editors, Mapping the Future of Biology: Evolving Concepts and Theories, Boston Studies in the Philosophy of Science, volume 266, Amsterdam: Springer, 2009, pp. 47-56.

3

between the two processes is clear in definition, it oftentimes collapses in practice.” (p. 168) Consider, again, the frictionless plane in which surfaces are idealized as perfectly smooth and collateral forces (friction, component forces, and so on) are suppressed. The same applies to population genetics, in which populations are idealized as infinitely large and factors such as random fluctuations are suppressed. These cases help elucidate the “in-tandem” character of the idealization/abstraction duality. Abstraction on a domain of enquiry D is sometimes a precondition of idealization over it. What is more, a theory’s domain may itself be an idealization of or abstraction from under-conceptualized “raw data”. Patrick Suppes has an attractive way of making this point. He reminds us that there are cases in which the data for a scientific theory must themselves be modelled. In those cases, then, what the theory supplies is a model of the models of those data.15

Notwithstanding ancient roots, contemporary excitement about models and their distortions may with convenience be dated from 1983, which saw the publication of Nancy Cartwright’s How the Laws of Physics Lie.16 Seen her way, some idealizations are of “properties that we conceive of as limiting cases, to which we can approach closer and closer in reality.” (p. 153) On the other hand, “some properties are not even approached in reality.” (p. 153) This is an important distinction. Some idealizations are real properties conceived of in the limit. Others are not real at all. Let us call the first group “within-limit” (WL) idealizations and the second “beyond-limit” (BL) idealizations. Then statements about WL-idealizations may be thought of as approximately true. Statements about BL-idealizations are unqualifiedly false, hence not in the least true. Both categories give rise to interesting philosophical questions, but BL-idealizations produce the greater challenge for the thesis that idealizations are cognitively virtuous. For how can the advancement of utter falsehoods ever rise to the status of knowledge?

2. Some motivating considerations

As we proceed, we want to be guided as much as possible by the reflectively intelligent practice of model-based science itself. This is where we will find our motivating data the pre-theoretical intuitions which we want our account to preserve and elucidate. Accordingly, we take it as given that

Science strives to attain a knowledge of things.

At its best, science succeeds in this task. In other words, instrumentalism in science is an unnecessary refuge.

Theories that advance our understanding of target phenomena possess cognitive significance.

Models of successful science are distortive.

15 Patrick Suppes, “Models of data”, in E. Nagel, P. Suppes and A. Tarski, editors Logic, Methodology and philosophy of Science: Proceedings of the International 1960 Congress, pp. 252-261, Stanford: Stanford University Press, 1962.16 Oxford: Oxford University Press.

4

Distortion is not incompatible with the attainment of scientific knowledge.

With these things said, a simple methodological protocol falls out: First, pay attention to the data. Second, do not abandon them without a fight.

3. Idealization In this section, our main target is the pure form of BL-idealization, the form of idealization to which there is no approximation in reality. As we said, examples abound: stars as perfect spheres, massless material points, perfect fluids, infinite populations, perfectly isolated systems, perfectly smooth surfaces, and so on. They bear no closeness whatever to anything in nature and, in relation to a theory’s target phenomena, statements about them are plain false; that is, false without an iota of truthlikeness. They are, as we might say, “extra-domain” with respect to those parts of the natural order that the theory in question seeks a knowledge of. For a significant range of cases at a certain level of generality, a scientific theory is a triple D, T, M. D is the domain of inquiry, and T the inquiry’s target with respect to D. Consider a case – typical of theoretical science – in which T is the specification of a mathematically formulated set M of lawlike connections. M we may take as a model of the theory in question. We say that a property is extra-domain iff it is essential to the theory of D but not present there, and that a property is intra-domain iff it is present in D. Similarly, a sentence is extra-domain iff it is ineliminably true in a model of the theory of D and yet false in D itself. A BL-idealization over D introduces a set of properties Q1, , Qn which cannot be instantiated in D but only in M. The Qi are true only of the objects in M, and assertions to the contrary are unqualifiedly false in D. Crucially, the lawlikeness of connections in M demand that the Qi occur essentially. BL-idealizations (both as properties and as attributions of those properties) are extra-domain.

We are, by the way, not unaware of the self-referential character of our present task. We are trying for a philosophical account of the cognitive good of saying what is false and suppressing what is true. Like any number of philosophical theories, there is nothing to prevent the account we develop here from appropriating the devices of idealization and/or abstraction. In other words, we want to leave it open that a good account of these features of model-based reasoning might itself be an instance of model-based reasoning. No doubt there will be readers whose methodological scrupulosity will resist modes of treatment of given target concepts which themselves instantiate those very concepts. Perhaps if we were seeking for simple lexical definitions of “idealization” and “abstraction”, a charge of circularity might be sustained. But since lexical biconditionals will bear little of the theoretical load here, we need not share – or welcome – our hypothetical objector’s methodological finickiness. The modelling of modelling may in some sense be circular but, if so, the circle would be virtuous, in the manner of Goodman.

There is an obvious question about an idealization’s Qi: Of what conceivable cognitive good are they? It is an unruly question. Since the dynamic equations of natural selection contain the Qi ineliminably, they are false “on the ground”. They are false, that is to say, of the actual populations in D. What is more, no population in D even begins to

5

approximate in size the populations of the model M. So there is no prospect of the truths about M’s populations being even truthlike about D’s populations.

All the same, the dynamic equations of M manage to do something quite extraordinary. They don’t tell us what populations on the ground are like, but they do play a role in telling us what natural selection on the ground is like.17

This what-likeness relation is crucial to our account. Here is why. There is in natural speech a well-entrenched use of “like”in which the what-like relation is irreflexive. In this usage, nothing can be like itself, not even exactly like itself. True, in formal treatments of likeness, reflexivity is often stipulated as a limiting case. But this just reinforces the point at hand. For if perfect self-likeness is only an ideal to which the likenessness of nature come more or less close, then the likenesses of nature – which are what natural speech is designed to capture – are irreflexive.

This has an interesting bearing on the role of idealizations in science. Given the irreflexivity of “like”, the respects in which x is like y are respects in which x is not. Accordingly,

FALSITY AND LIKENESS: There are ways of saying what’s false of x which reflect what y is like. (It is not ruled out that sometimes y = x). [EXAMPLE?]

Idealizing a thing is saying something false of it. Not all falsehoods are scientifically fruitful, needess to say. But when they are, they are likeness-exposing.

On reflection, perhaps this is no uncommon thing. There appear to be lots of cases in which a false description is genuinely illuminating; think here of the immense communicational value of metaphor. The point for our purposes is that knowing what things are like is a cognitive good, made so by the link between appreciating what things are like and having an understanding of them. Accordingly, getting population size wrong on purpose, and in a way that produces dynamic equations that are false in D, is also a means of achieving an understanding of natural selection on the ground. So we have it, as promised, that:

AID TO UNDERSTANDING: For significant ranges of cases, theoretical distortions of the sort examined here are indispensable to a scientific understanding of target phenomena.18

From which, also as promised:

COGNITIVE VIRTUOSITY: For significant ranges of cases, theoretical distortions of the sort examined here are cognitively (and not just instrumentally) virtuous.

17 Let us take some care with the “on the ground” metaphor. In its use here, what happens on the ground is what happens in a theory’s domain D, which, as we saw, may often itself be a model of the raw data.18 See here Batterman (2009): “ continuum idealizations are explanatorily ineliminable and a full understanding of certain physical phenomena cannot be obtained through completely detailed, non-idealized representations.” (p. 1)

6

This is interesting. It tells us something about the cognitive character of model-based science. It tells us that getting things wrong is often an unavoidable way of getting things right.

4. Side-bar: Classificatory confusion

The likeness thesis and its link to understanding put a certain pressure on how idealizations are to be described. Consider again the difference between BL (beyond-limit) and WL (within-limit) idealizations. Let Q be an idealized property of the BL kind (e.g. the infinite size of populations). Let us say that Q’s natural image is the size Q of an actual population at its physically biggest. In that case, the Q of Q-populations does not in any significant way reveal what Q-populations are like. Finite largeness is not like infinite largeness except in rather trivial ways. (For example, both are larger than smallness.) It is different with WL-idealizations. The athletic magnificence of an Achilles is only approximated to by the prowess of any actual athlete. A thinking agent who is smarter than any human actually is could be, but who is not infinitely smart, exhibits a smartness to which the smartnesses of nature clearly come more or less close. But all of them – the smartness of the super-smart agent and the various smartnesses of lesser beings bear the same closeness-relation to infinite smartness. They are all infinity and equally far from it. WL-idealizations, like finitely great smartness, have approximating natural images. BL-idealizations like infinite smartness have no natural images. There is nothing in nature that approximates to them. So it bears repeating that BL-idealizations don’t reveal what their natural images are like; but when properly wrought they can show what some other natural property is like (again, natural selection in finite populations).

Care should be taken to avoid a possible confusion. In the language of mathematical limits, finite quantities may approach infinity at the limit. But it is also true that for any arbitrarily selected pair of finite quantities neither is closer to the other to infinity; indeed neither is at all close to infinity. Accordingly, what is meant here by WL-idealizations are those on which a closeness relation is definable. What this helps us to see is that no idealization involving convergence on a mathematical limit at infinity is a WL-idealization in our sense. Part of the confusion lies in the failure – as in the case of Cartwright – to attend to relevant distinctions. There is a fair reading of Cartwright in which what we’re calling BL-idealizations are (a) properties not occurring in nature; (b) properties to which no nature-occurring property “comes close”; and (c) nature-occurring properties conceived of at their mathematical limit. What goes unremarked upon is not only that the three conditions are not equivalent, but they are actually inconsistent. The confusion deepens when we turn to WL-idealizations, which are characterized as (i) limit cases of properties occurring in nature; and (ii) properties to which nature-occurring properties come more or less close. But, as we have just seen, if the notion of limiting case is understood mathematically, no property satisfying (i) can ever satisfy (ii). Annoying as they might be, we needn’t be discouraged by these confusions. It suffices for our purposes that a WL-idealization is one to which a closeness-to relation is definable, and a BL-idealization is one for which no such relation exists.

5. Abstraction

7

It is now time to say something about abstraction. In its pure form, abstraction tells the truth incompletely and selectively. Abstraction serves as a kind of filter or to vary the figure partition. It partitions the truths at hand into those that matter in a certain way and those that don’t matter in that way. Thus what the filter captures is true, and what it leaves behind is also true. But these are truths with a difference. The former somehow count and the latter do not.

There exists in the literature no firmly established and precise unpacking of this notion of mattering. Certainly we know nothing in the literature that counts as a thick account of it. Distinctions have been made on the basis of several criteria. Among the more prominent we find relevance vs irrelevance, simplicity vs complexity, manageability vs unmanageability and tractability vs intractability ( McMullin 1985, Cartwright 1989, Weisberg 2008, Strevens 2008).19 To this we would add a further distinction which we’ll make some use of in sections 9 and following. It is the distinction between information and noise.

As before, D is a domain of enquiry, and T is the enquiry’s target with respect to D. Consider a case in which T is a mathematically formulated set M of lawlike connections. Here, too, M can be thought of a model. An abstraction AB is a filter on D which extracts real properties P1, , Pn featuring ineliminably in the lawlike connections worked up in M. Let us say that the elements P1, , Pn implicated in a lawlike connection are its parameters P1, , Pn, and that what an abstraction from D leaves behind is its residue R. Both the abstracted elements Pi and the residual elements Ri are intra-domain. They are all properties occurring in D.

Abstractions come in three different forms, depending on how the residual elements behave in relation to M. Although frequently confused in the literature, it is ruinous not to respect their differences. They are:

Faute de mieux (FDM) abstractions.

Essentiality (E) abstractions.

Simplicity (S) abstractions.

5.1 Faute de mieux-abstractions

We begin with the faute de mieux kind. There are classes of cases of which the following is true:

D = some set of natural phenomena including some event E.T = to attain a knowledge of E’s causes.M = the set of conditions causally necessary and sufficient for E.AB = an abstraction on D.Pi = abstracted properties from D occurring essentially in the lawlike connections of M.Ri = residual properties left behind – occurring in D but not in M and co-occurring in D with E.

19 For more, see the essays in Shanks (1998), and Jones and Cartwright (2005).

8

Think here of isolated systems, or motion in the absence of external forces.It is a point of considerable interest that for such cases, the residual Ri are such

that if admitted to M, the desired lawlike connections break down. In other words, the causal law for E is extra-domain. So we will say that

TARGET-BUSTING: Residues are target-busters.

Accordingly, it is a distinctive feature of FDM-abstraction that what is true in the model is false on the ground. For example, if M contains conditions causally necessary sufficient for E, they may be causally necessary for E in D, even though causally insufficient (which at least is something). Hence, they are faute de mieux.

We might note in passing that FDM models are plausible candidates for the role of partial structures and the support they lend to partial truths, in the manner of Newton da Costa and Steven French.20 Note, too, that FDM-abstraction resembles WL-idealization (via partial or approximate truth).

A good question is why we have faute de mieux models in the first place. If we could find necessary and sufficient conditions for E in D itself, we wouldn’t need M. We resort to M because it will give us what D itself won’t give us. And it might do so in ways that also give us some of what we were looking for in D.

5.2 Essentiality-abstractions

There are models aplenty in which lawlike connections are fashioned with causal minima, and are not subsequently broken when residual factors are added to the mix. Such models employ essentiality-abstraction. E-abstractions differ from FDM-abstractions in a quite particular way. They turn on a peculiar interplay of necessary and sufficient conditions. E-abstractions are compatible with the necessary and sufficient conditions on E affirmed in the model. They are also causally positive for E. But they are not causally necessary for E. Perhaps the most intuitive way of capturing what’s peculiar to E-abstraction is to represent its residual elements as causally redundant. They are, so to speak, causal surfeits. Accordingly, the function of E-abstraction is to model a causally redundant world irredundantly. What’s causally essential to E are its irredundant causes, and the inessential factors are the causally positive but redundant ones. E-abstraction captures something of Weisberg’s “minimalist idealization”, by means of which models capture “core causal factors which give rise to a phenomenon.” (2007, p. 642)

It is easy to see the duality between FDM- and E-abstraction. In the FDM-cases, residual elements are target-busters. When added to M lawlike connections break down. In the E-cases, the addition of residual elements to M leave lawlike connections intact.

5.3 Simplicity-abstractions

S-abstractions come in two varieties, both of which reduce the complexity of information. In one variation, their role is largely pedagogical. They filter complex

20 Science and Partial Truth: A Unitary Approach to Models and Scientific Reasoning. Amsterdam: Springer, 2005.

9

information into forms that can be managed by the comparatively untutored recipient. (In a variation of this variation, they serve as a handy kind of executive summary.) The second principal function of the S-abstraction covers two further kinds of case, one in which the original material, while comprehensible to the theorist, is costly to a fault to process. The other is one in which the complexity of a fully detailed account of a target phenomenon is incomprehensible, even to experts. Incomprehensibility, in turn, attaches to yet another pair of cases. In some situations, unfiltered explanations occasion informational overload, hence are too much for the theorist “to get his mind around”. In others, the informational surfeit is too much for the theory’s mathematics to “get itself around”. In the one case, we have a generalized inability to understand.21 In the other, we have a more focused failure of mathematical intelligibility.22

We see something of our S-abstractions in Strevens’ recent book in which he distinguishes between canonical versus textbook causal explanations. Canonical explanations are fully detailed accounts, which contain no idealizations, and omit nothing of causal salience to the targeted explanandum. Canonical explanations are not “better” than textbook explanations. In Strevens’ words, “an idealization does not assert, as it appears to, that some nonfactual factor is irrelevant to the explanandum; rather, it asserts that some actual factor is irrelevant.” (Strevens, 2008, p. 26) Strevens’ idealizations are our abstractions – a useful reminder of our earlier point about the absence of a fully settled usage of these expressions. They are filters of complex material whose particular purpose is to leave behind a residue of explanatorily irrelevant, unsalient, non-difference-makers. It is important to emphasize that what these residual elements fail to make a difference to is the theorist’s intended causal explanation. It is not Strevens’ contention that these factors are causally irrelevant; indeed the opposite is true. They are fully paid-up participants in the causal nexus on the ground. But they can be ignored without damage to the phenomenon’s causal explanation.

This idea of explanatory unnecessariness, that is, of explanatory non-difference-making, covers two kinds of situation, both catered for here. In the one, non-difference-making factors are unneeded for the construction of the causal connections embedded in a successful explanation. In the second, the non-difference-making factors aren’t necessary for the comprehension of the causal connections embedded in a successful explanation (and may be inimical to their comprehensibility.) In terms of the account developed here, Strevens’ “idealizations” that leave behind non-difference-makers in this first sense are essentiality-abstractions in our sense. His “idealizations” that leave behind non-difference-makers in this second sense are simplicity-abstractions in our sense.

S-abstractions also feature in Robert Batterman’s “On the explanatory role of mathematics in empirical science”, The British Journal for the Philosophy of Science

21 As mentioned earlier, the connection between models and understanding has been emphasized by Batterman. He cites a number of cases in which detailed models block the understanding. See Batterman (2009) and the further references cited there. Also of interest is Batterman’s “A modern (= victorian?) attitude towards scientific understanding”, The Monist, 83 (2000) 228-257, in which he draws a contrast between fully or highly detailed forms of explanation and understanding and their simplified forms which he himself favours. Writes Batterman: “In particular, the means by which physicists and applied mathematicians often arrive at understanding and physical insight involve a principled ignoring of the detailed mechanisms that form the essential core of the causal/mechanical views about understanding.” (p. 233).22 On guarding against the loss of mathematical tractability, see Weisberg (2007), p. 640.

10

(Batterman, forthcoming),23 in the distinction between “traditional” or Galilean idealizations (or abstractions, in our sense) and “non-traditional” or non-de-idealizable ones. Traditional idealizations are undisplacable through de-idealization; they contribute to explanation and understanding, “to the extent that they can be eliminated through further work that fills in the details ignored or distorted in the idealized models.” (p. 16) Non-traditional idealizations are not de-idealizable and are the more interesting type for Batterman, and for us. They play an essential ineliminable role in explanatory contexts (p. 17), by allowing the modeler to capture salient, robust features of target phenomena, which would lose their salience in a more detailed model. As Batterman says, “adding more details counts as explanatory noise – noise that often obscures or completely hides the features of interest.” (p. 17) As we will see, in due course, this is a point of crucial importance for our own account.

It is useful to point out that S-abstractions have residues that are target-busters, but with a difference. In the case of FDL-abstraction, its residual elements upon admittance to the model falsify the causal connections that obtain there. In the case of simplicity-abstraction, residual factors are compatible with laws of M and are part of the causal nexus for E in D. The mischief they do is not the loss of a causal law, but rather, in the one class of cases, the formulability of the law and, in the second, the comprehensibility of the law once formulated.

In relation to a model M formulating lawlike connections with respect to a D-element E, our abstraction trichotomy pivots on the quite different causal relations borne by residual factors to M or to E or to both. These can be summed up as follows:

With FDM-abstraction, the residual elements Ri override the lawlike connections that obtain in M and are sufficient conditions of their falsity in D.

With E-abstraction, the Ri are compatible with the laws of M, are causally positive for E but not causally necessary for it.

With S-abstraction, the Ri are compatible with the laws of M and necessary for E, but their recognition by M would render the desired account either unformulable or incomprehensible.

We note in passing the co-occurrability of S-abstraction with each of the other two. In the case of E-abstraction, it is perfectly possible that the elements they pick out as essential are also the very elements that make the abstraction a simplifying one, hence one that gives us a graspable understanding of E’s actual causal situation. Similarly for the case of FDM-abstraction. There are cases in which the construal of conditions which, as a matter of fact, are at best causally necessary for E as causally necessary and sufficient achieve a simplification that effectuates a partial understanding of E’s actual causal situation.24

23 Available online at doi: 10.1093/bjps/axp018.24 At the Model-Based Reasoning Conference in Campinas in 2009, Jaakko Hintikka proposed to us that no account of model-based science can pretend to completeness in the absence of an adequate treatment of the boundary conditions of differential equations in mathematical models. We agree with this entirely, and hope to make some progress with it in future work. The neglect of boundary conditions by philosophers of science is a regrettable omission. A significant exception is Mark Wilson’s “Law along the frontier:

11

When we introduced this trichotomy, we warned of the confusion caused by not respecting relevant differences. A case in point is the oft-repeated claim that the good of abstraction is a matter of its exposure of what’s essential and its simplification of what obtains on the ground. As we see, there are fair and load-bearing interpretations on which these claims are clearly inequivalent.

There is something rather intriguing about simplicity-abstraction. There are lots of cases in which there is more to E’s causal set-up than can be contained in any intelligible account of it. Yet grasping an abstract description of this same set-up suffices for an understanding of it. The residual elements of S-abstraction sometimes function as background conditions. Calling them this calls to mind the old mediaeval distinction between ordo cognescendi (the realm of knowledge) and ordo essendi (the realm of being).

With this distinction at hand, an interesting part of the story of simplicity-abstraction can be retold as follows: Since they are “background,” the elements Ri left behind by S-abstraction are causally distal in ordo cognescendi and causally proximate in ordo essendi. They are causally distal in ordo cognescendi because, even when we have the causal understanding of something furnished by an abstract description of it, we know in a non-individuating way that there is more to the causal picture than is captured by the abstraction. They are causally proximate in ordo essendi because they are causally active on the ground. This is rather important. If right, our knowledge of E’s causes is a distortion of E’s actual causal set-up. Why? Because the cognitive state one is when one has a causal understanding of E is a model of the actual causal situation that E is in. And, in so saying, the apparent innocuousness of the heuristic side of simplicity-abstraction is called into question. For how can it be right to say that one has a causal knowledge of E when one’s knowledge is a distortion of what’s really there – when ordo cognescendi is out of step with ordo essendi? Doesn’t having a knowledge of something require the smooth alignment of knowing and being?

The answer is No. Showing why is the business of Part II, immediately to follow.

II COGNITION

6. Cognitive systems

Our question now is: Why don’t the misalignments between ordo cognescendi and ordo essendi wrought by S-abstraction discredit their use in model-based science?

Differential equations and their boundary conditions”, PSA 1990, 2 (1991), 565-575 and Wandering Significance: An Essay on Conceptual Behavior, New York: Oxford University Press, 2006. Concerning the expression “boundary conditions”, Wilson writes “ standard philosophy of science texts encourage a rather misty and misleading understanding of this term. The faulty picture arises, as stereotypes often do, from inattention. Indeed, the standard philosophical texts say virtually nothing about boundary conditions – they are scarcely mentioned before they are packed off in an undifferentiated crate labelled “initial and boundary conditions” (usually pronounced as one word). The salient fact about “initialandboundaryconditions” is that, whatever else they might be, they are not laws and can be safely ignored.” (1991, p. 565)

12

The short answer is “Because this is the way of knowledge quite generally.” To see why, consider first some basic assumptions about the natural phenomenon of knowing things. In calling them to mind, we won’t be following the standard philosophical practice of “unpacking” the concept of knowledge. We propose instead to say something fairly simple and general about the admittedly complex business of how knowledge acquisition actually works in the lives of beings like us. In so doing, we won’t be checking the bona fides of these assumptions against their fit with this, that or the other pre-conceived philosophical account of knowledge.

Accordingly, we take it without further ado that the human individual is a cognitive system, that he and his ilk make their way through life by knowing things – by knowing what to believe and knowing what to do. We take it that the cognitive states in which a human individual finds himself is a state of nature and that, whatever the details, is the state which an agent is in when it is true of him that he has knowledge of something, a state that arises from the interplay of the agent’s cognitive devices and the causal surround in which they are activated. Cognition and its modes of production are, therefore, a kind of ecosystem in which cognitive targets are roughly proportional in ambition to the availability of resources sufficient for their attainment in more or less efficient ways and with a non-trivial frequency.25

Given the sundry distractions the human animal is heir to, it can only be said that, while perfection cannot be claimed, beings like us have done rather well in the cognitive arena. We survive and we prosper, and occasionally great civilizations arise. There are exceptions of course. There are cases galore in which we are mistaken in what we take for knowledge. Accordingly, there are two abundances that set out the basic contours of our cognitive lives.

COGNITIVE ABUNDANCE: Beings like us have knowledge, lots of it.

ERROR ABUNDANCE: Beings like us make errors, lots of them.

One of the striking things about the human individual is that error abundance doesn’t constrain cognitive abundance in ways that threaten survival and prosperity. From which we may conclude:

ENOUGH ALREADY: Human beings are right enough, enough of the time about enough of the right things to survive, prosper and occasionally build great civilizations.

How could this be? Why wouldn’t our being wrong about things on such a scale be enough to spoil our lives, if not simply wipe us out? The answer in part lies in the efficiency of our feedback mechanisms and the concomitant wherewithal for error-correction.

25 See H.A. Simon, Models of Man, New York: Wiley, 1957, Gerd Gigerenzer, Adaptive Thinking: Rationality in the Real World, New York: Oxford University Press, 2000, Dov M. Gabbay and John Woods, Agenda Relevance: A Study in Formal Pragmatics, Amsterdam: North-Holland, 2003, and Dov M. Gabbay and John Woods, The Reach of Abduction: Insight and Trial, Amsterdam: North-Holland, 2005.

13

Here is a last given. We take it that knowledge is valuable, that having it is indispensable to our survival and our well-being. If we were conspicuously less good at knowing things than we historically have been and presently are, life would be unpleasant or nonexistent. We all know Hobbes’ tag-line nasty, brutish and short.

There are legions of problems that arise from the relationship between a state of knowledge and what it is knowledge of. A great many of these fall under the umbrella of what Ralph Barton Perry called the “egocentric predicament”, which is a pessimistic reaction to the problem of the “veridicality” of direct experience. In one way or another, all these problems recognize the distortive circumstances in which cognitive states are formed, and virtually all of standard epistemology is a kind of answer to such questions. But this is not our focus here. Whatever might be the case – realism, idealism, Kantian transcendentalism or whatever else – we draw our line in the sand with the four assumptions currently at hand: cognitive abundance, error abundance, enough-already and the value-of-knowledge. There is one epistemological commonplace to which these assumptions give short shrift. It is scepticism. Scepticism is discouraged by the spirit of the cognitive abundance thesis and it is contradicted by the letter of the value-of-knowledge thesis when linked to the enough-already thesis. Beyond these modest musings, we will have as little to say of mainstream epistemology as we can get away with.

States of knowledge require information and states of information require causal provocation and causal support. In the interests of expository economy, we will focus on knowledge-states that can be considered both conscious and representational, but not before quickly adding that neither of these traits is a necessary condition on cognition.26

7. Knowledge as distortion

There are two general features of cognition on which we now want to focus:

A cognitive state is an abstraction of an information state.

Irrespective of its abstractive character, cognitive states are distortions of reality.

Beginning with distortion, you see a robin red-breast in the tree at the bottom of the garden. On the standard readings, the robin’s breast isn’t red. It is seen as red. This is Locke’s problem of “secondary qualities”, properties that are revealed in perception but are not part of nature’s inventory. Some philosophers extend the point to “primary qualities”, such as extension and magnitude. In a recent and extreme version, the denizens of the outer world are nought but “classes of quadruples of numbers according to an arbitrarily adopted system of coordinates”.27 Whether a secondary quality problem or also a primary quality problem, the veridicality issue is old-hat epistemology. It is known – or at least widely taken to be known – that the ways in which things are seen are

26 There is abundant literature on what Gabbay and Woods call cognition “down below”, that is, below the level of conscious awareness, some cases of which also fall short of the threshold of representational effect. See Agenda Relevance: A Study in Formal Pragmatics, Amsterdam: North-Holland, 2005, pp. 60 to 68 for a brief discussion and key citations.27 W.V. Quine, Theories and Things, Cambridge, MA: Harvard University Press, 1981; p. 17.

14

distortions of the causal stimuli of the perceptual states in question. Whatever the features of those perceptually provocative conditions on the ground, they aren’t red; and in some versions they aren’t breasted, and they aren’t robinesque.

Locke and a great many other philosophers (e.g. Hobbes and Condorcet) was of the view that the non-existence of colours precludes our knowing that the robin is red-breasted. So let’s ask: When you see the robin in the tree at the bottom of the garden, you are in a certain perceptual state. Is this a state that has cognitive value? Is it a state in virtue of which you know what’s at the bottom of the garden? Is it a state thanks to which you know what colour its breast is? If it is, then how can its cognitive virtue be denied? And if it is, we have significant ranges of cases in which knowledge of something is a distortion of its causally necessitating conditions.

Here is another case. The science of ornithology invests heavily in recording the distributions and conditions of change that affect the colour of birds. Non-trivial correlations are drawn between variations in colour and the presence or absence of other properties of significance. On the standard reading, ornithology is riddled with falsehood. Given the standard closure-conditions for conjunction, ornithology itself is false. So there is no knowledge of birds that can be got from ornithology. Aside from an occasional philosophical zealot, is there anyone in the wide world who believes such a thing?

Not everyone will like this position. How, it will be demanded, is it possible to know that the breast is red if there aren’t any colours? This is the wrong question. If you know that the robin’s breast is red, it isn’t true that nothing is red. What is true (as best we can yet tell) is that nothing red is causally implicated in knowing by looking that the robin’s breast is red. There are two things to say about this, both rather important. First, it is an interesting claim. And, second, even if false, it is not false enough, so to speak, to warrant outright dismissal. But it is at odds with a good many of the standard philosophical moves. This is because the standard philosophical moves assume that irrealism is true for colours and that irrealism precludes knowledge. It is precisely this that the present example rightly bids us to question. Everything of a scientific nature that is known of such things supports the distortion-as-a-cognitive-necessity thesis, and with it that the standard philosophical responses to irrealism eliminativism, scepticism, reductionism, idealism, instrumentalism, and so on are overreactions.

What this gets us to see is something quite important about the extra-domainality of features that are necessary for the advancement of D-knowledge. When we look at the bottom of the garden and conditions are right, we know by looking that the robin’s breast is red. The standard view is that since nothing in nature is red, the breast isn’t red and we can’t know otherwise. This is perverse. What is true is that the redness of the breast is extra-domain to the causal order of robins. Relative to that extra-domainality, robins themselves are causally intra-domain. Knowing by looking that the robin’s breast is red is a distortion of the causal conditions necessary and sufficient for that knowledge to exist.

8. Information suppression

We see the individual cognitive agent as a processor of information on the basis of which, among other things, he thinks and acts. Researchers interested in the behaviour of human information-processors tend to suppose that thinking and deliberate action are

15

modes of consciousness. Studies in information suggest a different view.28 Consciousness has a narrow bandwidth. It processes information very slowly. The rate of processing from the five senses combined – the sensorium, as the mediaevals used to say – is in the neighbourhood of 11 million bits per second. For any of those seconds, something fewer than 40 bits make their way into consciousness. Consciousness therefore is highly entropic, a thermodynamically costly state for a human system to be in. At any given time, there is an extraordinary quantity of information processed by the human system, which consciousness cannot gain access to. The bandwidth of language is even narrower than the bandwidth of sensation. A great deal of what we in some sense know – most in fact – we aren’t able to tell one another. Our sociolinguistic intercourse is a series of exchanges whose bandwidth is 16 bits per second.29

In pre- or subconscious states, human systems are awash in information. Consciousness serves as an aggressive suppressor of information, preserving radically small percentages of amounts available pre-consciously. To the extent that some of an individual’s thinking and decision-making is subconscious, it is necessary to postulate devices that avoid the noise, indeed the collapse, of information overload. Even at the conscious level, it is apparent that various constraints are at work to inhibit or prevent informational surfeit. The conscious human thinker and actor cannot have, and could not handle if he did have, information that significantly exceeded the limitations we have been discussing here.30

Here, then, is the basic picture. Knowledge is the fruit of information-processing. But it is also an information suppressor. There is a basic reason for this. Consider the particular case of conscious representational knowledge. If what is suppressed by our cognitive processes were admitted to consciousness and placed in the relevant representational state, it would overload awareness and crash the representation. From the point of view of cognitive attainment, what is suppressed would otherwise be experienced as noise.31 Noise, like bad money, drives out the good. It overloads perception and memory, and it hobbles conception. It blows the mind.

28 We are not unaware of the challenges that attend philosophical appropriations of the mathematical concept of information. However, for our purposes here we need not engage these issues. A good recent account is Fred Dretske, “Epistemology and information”, in Pieter Adriaans and Johan van Benthem, editors, Philosophy of Information, pages 29-48; a volume of the Handbook of the Philosophy of Science, edited by Dov M. Gabbay, Paul Thagard and John Woods, Amsterdam: North-Holland, 2008.29 M. Zimmerman, “The nervous system and the context of information theory”, in R.E. Schmidt and G. Thews, editors, Human Physiology, pp. 166-175, Berlin: Springer Verlag, 2nd edition, 1989. 30 Consciousness is a controversial matter in contemporary cognitive science. It is widely accepted that information carries negative entropy. Against this is the claim that the concept of information is used in ways that confuse the technical and common sense meanings of that word, and that talk of information’s negative entropy overlooks the fact that the systems to which thermodynamic principles apply with greatest sure-footedness are closed, and that human agents are not. The complaint against the over-liberal use of the concept of information, in which even physics is an information system,? is that it makes it impossible to explain the distinction between energy-to-energy transductions and energy-to-information transformations. See here, Stephen Wolfram, “Computer software in science and mathematics”, Scientific American, 251:188, September 1984. Difficult as these issues are, we say again that they needn’t be settled definitively for the concept of information to perform its intended role here.31 See, again, the noise-producing propensity of detail, noted in Batterman (forthcoming), p. 17.

16

If we had larger capacities, there would be less of this suppression of noise. For beings with brains “the size of a blimp,”32 what is noise for us would be serviceable information. But for us it is too hot to handle, and we must make do with states that filter out the noise that would otherwise overwhelm them. This supports the abstraction thesis: A cognitive state is an abstraction from an information state. There is another way of saying the same thing: A cognitive state is a model of an informational environment.

9. Connection to models

In what we have said so far, perhaps the connection with models is not wholly apparent. As with knowledge, a scientific model is a certain kind of information state. There are cases galore in which if abstractive-residues were admitted to them, they would lose their informational purchase. Information contained in the residue would operate as noise in the model. So the model excludes it. It is much the same way with cognition in the general case. A piece of knowledge is an information state meeting certain conditions. It is a state that filters manageable from unmanageable information. In the interesting cases, excluded information is the residue of abstractions; we might think of these as informational surrounds. Informational surrounds have some of the interesting properties we have been tracking here. If admissible, their elements R1, Rn would make the knowing agent better informed about the thing at hand. But if actually admitted they would crash the host state. Informational surrounds are good information which the knowing agent cannot make use of – cannot get his head around. So when they are in good working order, his cognitive devices keep at bay the Ri of these informational surrounds.

Informational surrounds bear a certain likeness to background knowledge. Perhaps their most interesting point of similarity is that, while inaccessible to consciousness, informational surrounds and background information are causally implicated in the attainment of those informational states which, by wide agreement, qualify as states of knowledge, including the conscious ones. Thus when, looking into the garden, you come to know that there is a robin there, much of the information contained in your visual field is screened out. But it is screened out in a quite particular way. On the one hand, it is not part of the informational state that constitutes knowing the robin is there, but may well be causally necessary all the same to the creation of that state. One of the problems that launched us into these reflections arises from the fact that often a residue (or causal field) if admitted to a model that excludes it would blow the model, and hence is a defeater of the model’s lawlike pronouncements on the ground.

We wanted to know what was the cognitive good of models that exclude falsifying considerations in this manner. What we now propose is that scientific models mimic the structures and processes of cognition quite generally, and that a fully worked up model-based scientific theory would capture with some precision the constructive impossibility of knowing a thing through an awareness of most of what’s true of it. With perception again as our guide, knowing of the world what you do when you see the robin in the tree is, in comparative terms, knowing hardly anything that’s true of it. Such knowledge – a conscious awareness of the disclosures of your five senses – beckons paradoxically. It supplies you with ample knowledge on practically no information. It is

32 Stephen Stich, The Fragmentation of Reason, Cambridge, MA: MIT Press, 1990.

17

not that the screened-out information is suppressed because of its uninformativeness about what you now perceive. It fairly brims with information about those objects. It is excluded because you can’t process it, can’t take it in. It is excluded because, if admitted, it would erase the little that you already do in fact know of it.

In these respects, the abstractions of model-based science mimic the abstractive character of knowledge in general. If, as we ourselves suppose, the abstractive character of perceptual states doesn’t deny them cognitive value, why would the abstractive character of model-based reasoning deny it cognitive value? After all, shouldn’t it be that what’s sauce for the goose is sauce for the gander?

III ARTEFACTUALISM

10. Truth-making

In the case of perceptual knowledge we have it thus. We see the redness of robins. We know that robins are red-breasted. It is true that robins are red-breasted. Since knowing is achieved by being in states which fulfill the relevant conditions, it is natural to ask what role redness has in producing the state that meets the conditions for knowing that the robin is red-breasted. The answer is none. Although it is true that the breast is red, although we know that it is, it is not true that redness makes any appearance in the causal arrangements necessary for that knowledge.

From the point of view of models, we can now say that

PERCEPTIONS AS MODELS: In certain respects, a field of vision is a model of its causal surround, both of which are states of nature.

In contrast, consider once more the case of population biology. We have long since settled it that infinite populations, which don’t occur in nature, are necessary for a knowledge of natural selection, which does occur in nature. Since infinite populations don’t occur in nature, they have no occurrence in the causal mix without which natural selection on the ground could not occur. Whatever the precise nature of the tie between infinite populations and the knowledge attained of natural selection, it is something different from their occurrence in the underlying causal surround.

Although the red-breastness of robins is extra-domain with regard to its causal surround, both the red breastness and the causal surround occur in nature. Similarly, although our knowledge of the red-breastness is extra-domain with respect to our knowledge of what causes the knowledge of red-breastness, both pieces of knowledge are true of the natural world. We know in each case where these things are true.

The infinite populations of population biology are different. Although infinite populations are extra-domain with respect to natural selection, natural selection occurs in nature but infinite populations do not. Similarly, although our knowledge of infinite populations is extra-domain to our knowledge of the processes of natural selection, what we know of natural selection is true of the natural world, whereas infinite populations are not true of it. We know in the one case where it is true, and in the other also know that it is not true where the former is. Accordingly, if we wish to retain the commonplace that only the true is known, there is a question that now presses:

18

TRUTH PLACES: Where, if anywhere, is it true that populations are infinite? What, so to speak, is the truth-place of “Populations are infinite”?

This is a philosophical question. It arises, quite properly, from the claims we have been making on behalf of the cognitive virtuosity of distortion. We said in section 1 that in this paper we would give short shrift to the philosophical questions provoked by what we would come to say of distortion; that beyond a short sketch of possible answers, we would be content to offer readers a promissory note, to be redeemed in future work.33 Here, then, is a short sketch of an answer to the home-truth question for BL-idealizations. It is a version of artefactualism known as “stipulationism”.

In its most general form, artefactualism is the doctrine that some things are made rather than found and that their main modes of production are either the linguistic acts or the cooperative behaviour of human agents. The two main varieties of antifactualism are conventionalism and stipulationism. Conventionalism is the view that some facts are constituted by the cooperative behaviour of agents, and that sentences reporting these facts are also made true by this behaviour. Think here of the fact that in North America and Europe the right side of the road for motorists to drive on is the right-hand side. Stipulationism is the view that some facts are constituted and some statements about them made true by the linguistic acts of agents. Think here of the fact of legal guilt, constituted by a jury’s assertion of it, or the fact that Parliament is now in session, made so by the Sovereign’s proclamation. Stipulation is the form of artefactualism that matters most for our purposes here. So we won’t trouble ourselves further with conventionalism.

10.1 Stipulationism

Stipulationism, then, is the view that some objects and some states of affairs are made rather than found, and that involved in their making is their being thought up. Stipulation likewise provides that objects and states of affairs are thought up in ways that make sentences about them true. Made objects make true the sentences that faithfully announce their presence and faithfully report their goings on. They are made, and made true, by fiat. Fiat is Latin for “Let it be made.” “Let it be made” is a sentence in the hortatory subjunctive mood. This tells us something important about making. It tells us that artefactualism’s objects and truths are, once they are thought up, uttered into being; they are, as we have said, the results of their makers’ linguistic acts.

There are critical differences for the artefactualist to take note of. One’s knowledge by looking that robins are red-breasted depends on our seeing the redness of the relevant breasts, and is the result of causal provocations from which redness is missing. It is true that robins are red-breasted because it is known that they are. A scientist’s models are different. Their fruitfulness or otherwise for scientific enquiry has nothing directly to do with whatever may be the causal conditions of the theorist’s thinking them up. It is true in nature, but not in the relevant causal nexus, that robins are red-breasted, and it is not true in nature that a population is infinite. If parity with the robin-example is to be got, we must find a sense of “true” and a place in which it is true that populations are infinite. If we succeed in this, the place in which the infinite

33 For example, in “Model-based realism”, to appear.

19

population statement holds will be extra-domain with respect to the place in which the natural selection statements are intra-domain.

Where is this somewhere? According to our artefactualism, it is in the model of the theorist’s theory. What makes it true there? From everything that is known of the natural history of scientific (and mathematical) enquiry, the answer is “By fiat”. Consider the cases: let there be an x such that ; we put it that ; we define a function f over ; and so on. Artefactualists say that these sentences are made true by conditions other than designatory links to elements of the natural order. Stipulation answers the question: “How then are they made true.?” The answer is that they are made true by the theorist’s stipulation. Stipulation has a long history in modern philosophy, and has played long since a dominant role in mathematics.34 Stipulations are a good deal more widely embodied in enquiry than are idealizations and abstractions. When a mathematician stipulates a function f over a set that meets certain conditions, he needn’t think that f is either an abstraction or an idealization in the scientist’s senses of these terms. We ourselves are drawn to stipulation for two particular reasons. One is that is there. That is, it is an established part of scientific – especially mathematical – practice, and not an ad hoc device of our own contrivance. (In other words, stipulation is not the product of our own stipulation.) The other is that it helps with the task of finding a place in which extra-D claims essential to a scientific knowledge of D can turn out true. In short, it helps in finding a truth-place for them.

Stipulationism is an answer to the question of how something false in D can be indispensable to a knowledge of D’s phenomena. It provides that it really is the case that sentences false in D are nevertheless true, in the absence of which certain things known of D couldn’t have been known. The central question to which this gives rise is: How is it possible for the theorist’s stipulations – his sayso to generate truths?

10.2 Fictionalism

This too is a large question, for whose honest consideration we lack the space here.35 Suffice it to note a burgeoning literature which explores an interesting and attractive suggestion. It is that the sense in which sentences false on the ground can be true in a theory’s model resembles the sense in which sentences false on the ground can be true in a novel or short story. It is a controversial issue, in which various issues of greater and less importance remain unsettled. But, for good or ill, the idea is now making

34 For Kant, stipulation is synthesis, which is the making up of new concepts. For Russell, stipulation is mathematical definition, which makes things true without the generation of the concepts in virtue of which this is so. See Immanuel Kant, Inquiry Concerning the Distinctness of Natural Philosophy and Morality, Indianapolis: Bobbs-Merrill, 1974; first published in German in 1764, and Logic, Indianapolis: Bobbs-Merrill, 1974; first published in German in 1800; and Bertrand Russell, Principles of Mathematics, London: Allen and Unwin, 1937, p. 27; first published in 1903.35 Interested readers may wish to consult our “Unifying fictions”, in Fictions and Models: New Essays, which also contains a foreword by Nancy Cartwright and essays by Robert Howell, Amie Thomasson, Mark Balaguer, Otávio Bueno, Mauricio Suarez, Roman Frigg, Jody Azzouni, Alexis Burgess and Giovanni Tuzet. See also John Woods and Jillian Isenberg, “Psychologizing the semantics of fiction”, forthcoming in Methodos, and Woods’ Préface, in Fiction: Logiques, Langages, Mondes, Brian Hill et al., editors, forthcoming in 2010 in the series Les Cahiers de Logiques et d’Epistemologie from College Publications, London.

20

the rounds in mainstream philosophy of science that the facts reported by true theories indispensable to a knowledge of D that have no possibility of occurrence in D are fictions. Since fictional sentences are true in their place that is, stories are the truth-places for them it is said that this gives the desired result, namely, a non-trivial sense of irrealism in science which evades the drab consolation prizes of idealism, reductionism, instrumentalism, and constructive empiricism, to say nothing of the sheer capitulations of eliminativism and scepticism.

Concerning literary fictions, a standing position is not only that there is an obvious sense in which they fail realist conditions, but that they fail them on purpose. No one – not even a literary creator – thinks that the objects and goings-on of fiction are real, or that their faithful reports are true of reality. On the other hand, it is just about as widely held that there is a place in which these objects and events do occur and of which the sentences that faithfully report them are true. One of the main tasks of a theory of literary fictions is to supply a theory of truth that accommodates this truth-place approach. It tries to specify conditions under which sentences false (or not true) in the real world are true in the story. It is not an easy task, and we are a long way from a settled theoretical consensus. Even so, there are points of wide agreement about particular matters. One is that authors make true the sentences of the text of the story. Another is that, except where borrowed from the world, the story’s truths are false (or at least not true) on the ground. A third is that, subject to conditions it is difficult to specify, fictional truths semantically engage with non-fictional truths to produce further truths. (For example, “Sherlock Holmes had tea with Dr. Watson” which is true in the story and, “Having tea with is a symmetrical relation” which is true in the world together imply “Dr. Watson had tea with Holmes”, which is true in the story.

Stipulationists assert a parity with the creative efforts of model-based science. Accordingly, theorists make sentences true in their models. These sentences are false in the world, or more generally, in the theory’s domain of enquiry. Nevertheless, sentences true in the model also semantically engage with sentences true in the domain to produce further sentences that are true in the domain.36 There is substantial agreement that a principal task for a theory of fiction is to produce a theory that takes proper account of the three facts noted in the paragraph just above. And there is growing agreement that a principal task for a stipulationist theory of models is that it take proper account of the similar-seeming trio of facts noted in the present paragraphs. If these tasks were to be successfully discharged, then we would have a principled account of truth-places for fiction. We would also have a principled account of truth-places for model-based science.

Fictionalism in science is the view that the success-conditions for a theory of truth-places for fiction and the success-conditions for a theory of truth-places for science reflect a notion of fictionality which both those theories instantiate. Currently in question is whether the common notion is the literary one adapted to scientific contexts,37 or whether the common concept is something more generic than that.

36 But let us note in passing what appears to be a significant difference between the two cases. In the scientific case, fictional truths can mingle with real-world truths to generate new real-world truths. In literature, it is the other way around. Fictional truths semantically engage with real-world truths to generate further fictional truths.37 For the pro case, see e.g. Roman Frigg, “Fictions in science”, in Fictions and Models: New Essays. And for the con case, see again our own “Unifying the fictional”.

21

Of course, it hardly needs saying that not all stipulationists, that is, all who accept the above three claims – true in the model; false in the domain; yet subject to semantic integration – are happy with a fictionalist account of them. This, too, is an open question in the relevant research communities.

A theory of truth-places carries certain suggestions, all of them alarming to philosophers with pre-conceived leanings. One is that “true” is at least a two-place, rather than unary, predicate. Another, relatedly, is that truth is subject to a relativism of place. A third, again relatedly, is that truth-places commit the theory of truth to a kind of semantic pluralism. A fourth is that a theory of truth-places doesn’t rule the consequence, and is not intended to, that good scientific theories can flourish notwithstanding – indeed on account of – sentences false in the theory’s domain, hence the negations of sentences true in it.

These are consequences that shock orthodox opinion. Clearly something has to give. Either truth-places must go, or orthodoxies must be relaxed. This sets up two options. The consequences must be suppressed, and with them the renegade theory which gave rise to them. The other is to regard the consequences as constituting new research options, new projects for enquiring minds. As of this writing, these are projects already well-underway. Concerning the binary character of the truth predicate, its relativity to places and the promptings it gives to pluralism, there is now a brisk trade in logical pluralism, and an equally brisk one in truth-in-fiction.38 And concerning the inconsistencies engendered by the domain-falsity of model truths, there are prosperous and beckoning projects in paraconsistent logics, which are logics that effectively de-claw the classical menance of inconsistency.39

We said earlier that to the extent possible we want to preserve the data to which truth-places are a measured response. Orthodox capitulation marks an end to this fidelity; it sends the data packing. The research programmes just mentioned, and others underway or in prospect, offer the data a reprieve. In our view, this is the way to go. As we said, we are not prepared to surrender our data in the absence of a hard case to the contrary. But since the orthodoxies against which our data rub were not fashioned in reaction to them, giving them standing now would be to violate the methodological requirement to pay 38 For the first, see for example, John Woods and Bryson Brown, editors, Logical Consequence: Rival Approaches, Oxford: Hermes Science, 2001, John Woods, Paradox and Paraconsistency: Conflict Resolution in the Abstract Sciences, Cambridge: Cambridge University Press, 2003, and JC Beall and Greg Restall, Logical Pluralism, New York: Oxford University Press, 2006. For the second (and for an extensive bibliography) see Woods, “Fictions and their logic”, in Dale Jacquette, editor, Philosophy of Logic, pp. 835-900, a volume of the Handbook of the Philosophy of Science, edited by Dov M. Gabbay, Paul Thagard and John Woods, Amsterdam: North-Holland, 2005.39 See, for example, Newton da Costa, “On the theory of inconsistent formal systems”, Notre Dame Journal of Formal Logic, XV (1974) 497-510; Dederik Batens, “Paraconsistent extensional propositional logic”, Logique et Analyse 23 (1980), 195-234; Graham Priest, Richard Routley and Jean Norman, editors, Paraconsistent Logic, Munich: Philosophia Verlag, 1989; Dov M. Gabbay and Anthony Hunter, “Making inconsistency respectable”, part 1, in P. Jourand and J. Keleman, editors, Fundamentals of Artificial Intelligence Research, 535 (1993), pages 19-32, Berlin: Springer Verlag, 1991; and part 2 in LNCS 747, pages 129-136. Berlin: Springer-Verlag, 1993; Newton da Costa and Otávio Bueno, “Consistency, paraconsistency and truth”, Ideas of Valores, 100 (1996), 48-60, Woods, Paradox and Paraconsistency, 2003, Da Costa and French, Science and Partial Truth, 2005, Woods, “Dialectical considerations on the logic of contradiction I”, Logic Journal of the IGPL, 13 (2005), 231-260, and Peter Schotch, Bryson Brown and Raymond Jennings editors, On Preserving: Essays on Preservationism and Paraconsistent Logic, Toronto: University of Toronto Press, 2009.

22

them due attention. So we will stick with truth-places until such time as arguments tailor-made against them force us to stand down.

11. Irrealism without tears

One of the morals suggested by the preceding pages is that, as standardly taken, the realist/irrealist divide is a piece of antique taxonomy that lags behind the relevant facts. Properly understood, realism-irrealism is not a partition. There are philosophical appreciations that fall between the two classical extremes and their shared presumption that winners take all. If the claims that we have advanced here can be made to stand, then there is a perfectly good sense of “irrealism” and a perfectly good sense of “true” according to which some irrealist theories give us a knowledge of D-phenomena by way of truths that hold not in D but only in the theorists’ model M. And it is a sense of “irrealism” for which none of the classical options to realism is supportable, never mind necessary. If so it breathes new life into an old distinction. In its refreshed form, the distinction is relativized to truth-places. Colour-recognizing theories are realist as regards the place in nature in which colour-statements are true and irrealist as regards the place in nature in which the causes of colour-perception exercise their powers. Idealized and abstractive science are realist with respect to the places they are true (namely in models) and are often irrealist with respect to which their target phenomena occur (namely in nature). The concept of truth-places accommodates the extra- and intra-domain distinction in the obvious way. And it helps us see what is surely present in actual scientific practice, namely, that since parts of theories can be true (and false) in different places, they can be both realist and irrealist – irrealist in relation to the places where their claims are false and realist in relation to the places where they are true. It is perfectly true that there are monomaniacal monists lurking about, according to whom valid science must recognize only one place for the true. There’s no accounting for philosophical tastes it seems, but nothing in reflectively intelligent scientific practice lends any encouragement to this vaunting monoplism.40

Pages ago, we said that to the extent possible we would try for a thick account of the cognitive virtuosity of model-based distortion without presupposing this or that pre-set philosophical position on knowledge. We bring this essay to a close with a further few remarks on this same theme. At a certain level of generality, most by far of the going positions in epistemology all cleave to a common notion about knowledge. They see knowledge as discovery. Thus to come to know that φ you discover φ’s truth. The discovery paradigm cuts no slack to embedded practices in mathematics and theoretical science. These are the practices of stipulation, of making φ true rather than discovering it. Clinging to the discovery paradigm carries hefty philosophical costs for what in many ways are the paradigm cases of knowledge indeed are the best ways of knowing things. An inflexible attachment to the discovery paradigm infests the mathematical and theoretical sciences with irrealism, and with it one or other of the set-piece reactions to it. We said at the beginning that we weren’t going to stand still for this, that we weren’t prepared to countenance the nonsense that our best ways of knowing things fail to produce knowledge. Any philosophical theory requiring us to concede that we don’t have

40 For a detailed exploration of this theme, see again our “Model-based realism”, to appear.

23

a knowledge of natural selection on the ground deserves outright rejection. But it is a rejection with consequences. One of them is that the discovery-paradigm can’t be right for all of knowledge.

This is the right thing to reject. Although one way of knowing something is by discovering what’s true of it, another is by making things true of it. This is knowledge, not by discovery, but by creation; and it calls to mind Quine’s jape that the knowledge conveyed by theories is “free for the thinking up”, and Eddington’s that it is a “put-up job.” What is so striking about created knowledge is not so much that it lies within our power to make, but that it integrates so robustly with discovered knowledge. This – the phenomenon of semantic engagement – achieves a deep grip. For there are cases in which discovered knowledge is impossible in the absence of created knowledge. This is the cognitive message of model-based science.41

John Woods Alirio RosalesUniversity of British Columbia University of British Columbiajhwoods@interchange.ubc.ca arosales@interchange.ubc.cawww.johnwoods.ca

41 Acknowledgements. The first version of this paper was presented in June 2007 to the Canadian Society for the History and Philosophy of Science in Vancouver. For some helpful prodding, both via voce and epistolary, we extend our thanks to Anjan Chakravartty, Michael Weisberg, and most especially Michael Strevens. The present version was given as an invited lecture at the MBR09, a conference on model-based reasoning, meeting in December 2009 in Campinas. For welcome support and instructive criticisms we are indebted and grateful to Lorenzo Magnani and Walter Carnielli (the conference organizers), Jaakko Hintikka, Paul Thagard, Balakrishnan Chandrasedaran, Peter Bruza, and Cameron Shelley. For correspondence or conversation on the role of models in science, we also thank Patrick Suppes, Bas van Fraassen, Steven French, Margaret Morrison, Dov Gabbay, Otávio Bueno, Roman Frigg, Sally Otto, Michael Whitlock, Mauricio Suarez, and Shahid Rahman.

24